The Trump administration is preparing a new voluntary program to test some of the nation’s most powerful artificial intelligence models for dangerous cybersecurity capabilities, including their ability to find software vulnerabilities and carry out sophisticated computer-based tasks with limited human supervision.
The plan came into sharper focus after White House officials met Tuesday with representatives from Meta, Anthropic, Google, Nvidia and OpenAI. Under the approach discussed at the meeting, the government would focus its testing on powerful closed AI models while leaving open-weight models outside the program.
Closed models — including leading systems developed by companies such as OpenAI, Google and Anthropic — remain controlled by their developers. Open-weight models, by contrast, make their underlying model weights available for others to download, modify and deploy. The distinction is significant because the administration has made encouraging open AI development part of its broader technology strategy.
The cybersecurity testing effort comes amid growing concern about what increasingly autonomous AI systems can do when given access to computers and online tools. OpenAI and Anthropic have recently disclosed incidents involving AI agents that accessed outside computer systems during security evaluations, intensifying scrutiny of models capable of finding vulnerabilities and performing sophisticated cybersecurity tasks.
The testing framework remains voluntary rather than a government approval system for new AI products. But Tuesday’s meeting clarifies an important part of Washington’s approach: the government intends to scrutinize some of America’s most capable proprietary AI systems while, for now, taking a different approach toward models whose underlying weights are openly available.
